A Spectrum Sensing Algorithm based on distributed cognitive models

Andrea F. Cattoni*, Irene Minetti*, Matteo Gandetto*, Ruixin Niu°, Pramod K. Varshney°, Carlo S. Regazzoni* * Department of Biophysical and Electronic Engineering (DIBE) - University of Genova {cattoni, irene, gandetto, carlo}@dibe.unige.it ° Engineering and Computer Science Department (ECS) - Link Hall - Syracuse University {rniu, varshney}@ecs.syr.edu
single CT will be defined. In the last paragraph, the usefulness of CMs is proposed as a basis to obtain more complex cooperation strategies among multiple CTs operating in the same environment; a simulative framework will be introduced to present results showing the validity of the proposed approach. 2. COGNITIVE RADIO TECHNOLOGY The idea of Cognitive Radio, as a new approach for wireless communication was first presented by Joseph Mitola III [1]. It was thought as the final point of evolution for a software-defined radio platform, considered now as a black-box that changes its communication functions depending on network and/or user's requirements. After Mitola, other Researchers gave their definitions, as Bruce Fette [2] or Simon Haykin [3]. Also governmental agencies have tried to provide a standard definition for CRs, clarifying how this technology can be used too. One of the first document which deals these problems is the Notice of Proposed Rule Making and Order (NPRM), compiled by the U.S. Federal Communication Commission (FCC) [4]. In this document, the agency confirm that CR technologies can allow a more efficient and dynamic spectrum usage, self-localization of terminals, frequency, modulation and transmitted power allocation. Summing up, it’s possible to affirm that the CRs are adaptive radio terminals, aware of their potentialities, of the environment and of their target, and they are able to learn, for example, through the experience new waveforms, new environmental models and new operative frameworks. In this paper, attention is focused in the issue of representing information acquired from radio channels into appropriate environmental models that are shared by multiple CTs in order to accomplish radio context discovery tasks. Even though represented environmental Cognitive Maps are in this paper considered as a-priori available from CTs, the capability of incrementally learning such maps can be an extension of this work. The presented approach shows how each CT can take advantage of simple existence of other CTs in the same environmental area covered by the shared CM to improve its efficiency and robustness in making simple decisions. The problem of identifying transmission modes of heterogeneous

ABSTRACT In the last years, an increasing attention of the communications researchers has been focused on the Cognitive Radio (CR) concept and on its possible supporting technologies and applications. The proposed approach deals the problem of information acquisition and handling for cooperative Cognitive Radio Terminals, in order to perform Spectrum Sensing tasks in a distributed way. The proposed solution is based on Distributed Detection theory supported by a Cognitive Modeling of Terminals. 1. INTRODUCTION In the last years, an increasing attention of the communications researchers has been focused on the Cognitive Radio (CR) concept and on its possible supporting technologies and applications. CR is considered as a new goal for wireless communications in the next future. This kind of technology will allow a more efficient and flexible usage not only of the electro-magnetic spectrum, but also of all the resources of the Cognitive device, like, for example, batteries and computational capacity. The capability of observing the surrounding environment, by understanding the characteristics of the operative context and by adapting their operations to it in a completely autonomous and intelligent way, will be the peculiarities of the Cognitive Terminals (CTs). CRs can be considered as the natural evolution of Software Radios (SRs): starting from the flexible and completely digital physical architecture of SRs, the goal is to design intelligent terminals with the previously described characteristics. Different disciplines are involved in the design and development processes of CTs, from electronics to bio-information sciences, from signal processing to telecommunication engineering: all of them are required for the realization of all the required components of CTs, both on algorithmic and hardware platforms sides. In this paper, a brief definition of CRs will be first provided, together with a proposed model for the description of the operating strategy a CT should follow. Then, the concept of Cognitive Maps (CMs) is discussed and their role in the behavioral model of a

Proceeding of the SDR 06 Technical Conference and Product Exposition. Copyright © 2006 SDR Forum. All Rights Reserved

even if these works are not explicitly focused on Cognitive Systems. but it isn’t able to provide which standard is occupying the examined bandwidth.2. In Figure 1 a representation of a Cognitive Cycle is presented. In the state of the art. In the second step (Analysis) the acquired data are processed and analyzed in order to provide to the system an higher level synthetic representation of the context. within the Cognitive Cycle. but to create a network of cooperative terminals. because they are able to extract proper features and to perform a suitable classification of the modes. identifying if the bandwidth is already occupied by a transmitted signal. which can cooperate in order to obtain better performances. All Rights Reserved . In fact different methods [7][8]. Hence spring out the need of maps which allow the terminal to orientate itself within the context. DISTRIBUTED MODE IDENTIFICATION The Spectrum Sensing and Mode Identification (MISM) process plays a key role in the Cognitive Radios because it provides an observation of the physical world: this is the knowledge about the channel conditions which allows to take a proper decision for the current context. The knowledge about the physical world can be organized. Analysis. This feature grew in importance in the last years. The Decision Map contains information about which is the decisional process which has to be performed for a proper reaction. influences the behaviour of the terminal through variations of this internal representation. into four maps. This approach is characterized by a very low computational load. but with an explicit relationship with the phisicity of the problem. to be able to interact with it. in order to perform a fast and optimal re-configuration of the intelligent device. The action represents an active interaction with the external environment because the CT tries to influence the other interacting entities in order to maximize its internal functional. The oldest and simplest one is the Radiometer [6]: it extracts the energy in each sub-band. it interacts with the external world through four main steps: Sensing. Decision and Action. It is also shown that distributed decision theory [5] can be used as a way to exploit shared CMs knowledge within a well-assessed probabilistic distributed decision problem. 3. The basis for the proposed approach can be found in [15]: the same operative framework is considered consisting in an indoor environment with two sources Figure 1 Simplified Cognitive Cycle The Cognitive Cycle occurs in a continuous way and during the entire process the CT uses the observations and its decisions in order to improve dynamically its behavior: it can be considered as a continuous learning phase. which face also the problem of identification of modes superimposed on the same bandwidth [9][10]. Copyright © 2006 SDR Forum. The first map. an internal representation of the physical world. while the Action Map describe how to carry it out. Besides. Cognitive Systems are thought to be used not only in stand alone mode. the Sensing Map provides information on how to observe the environment in each point of the space. 3.wireless sources of known position is considered here as a case study. These methods. The four main steps are here clearly evidenced. as all the previous algorithms work. The Cognitive Cycle The Cognitive Cycle is a model which describe the behavior of any living being. In fact. the choice is based on the embedded internal knowledge. Organizing the internal knowledge into maps has manifold advantages: first of all it allows to describe the current context in a way understandable by the system. by creating a direct relationship between the physical world an the inner representation. the past experience and the current context. this final map is the interactional interface between the terminal and the physical world. A first step into this direction. Map Concept Any Cognitive entity which is immersed in a physical environment needs. different implementations for spectrum sensing are already present. one for each stage of the cycle. The first stage of the cycle (Sensing or Observation) represents a passive interaction of the terminal with the environment: the CT gather information about both its internal state and the surrounding environment in a continuous way.1. The Analysis Map allows the Cognitive Entity to reach a semantic and contextual representation of the external environment. have been proposed. in framework of Cognitive Radios. can be found in [13][14]. these works face also the problem of distributed detection under communication [13] or energy constrains [14]. In the Decision stage the Cognitive system has to decide which is the most proper action to the received external stimulus. COGNITIVE MODELING 3. acquired with the experiences. It’s also easier to exchange knowledge between multiple terminals. involve also the Analysis stage. The knowledge. Proceeding of the SDR 06 Technical Conference and Product Exposition. 4. it’s possible to improve the adaptivity to the external conditions.

• Only Bluetooth source is transmitting. By considering similar the “point of view” of the two terminals. Copyright © 2006 SDR Forum. that are not considered yet in the paper. The peculiarity of the problem consists in the bandwidth overlapping (both are in the ISM band) of the two standards. A 12x12 meters room. Feature extraction and Mode Classification modules represent instead the Analysis module. of the signal components [15]. Four possible transmission situations can be present: • Only WLAN source is transmitting. All Rights Reserved . 5. they represent the Sensing module of the CCR device. able to move inside the room. the problem is how to use it. but. WV transform has been used in order to exploit the signal characteristics intrinsic in the two transmission modes: in Figure 3 the different time-frequency behaviors for the two signals are shown. It’s hence possible to jump to the second stage of the Cognitive Cycle. Proceeding of the SDR 06 Technical Conference and Product Exposition. • Both sources are switched off. inside a time window.transmitting WLAN or Bluetooth.1. ω ) = 1 2π ∫S * (ω + 1 θ ) S (ω − 1 θ )e − jtθ dθ 2 2 Goal of the CTs in the considered framework. containing a Bluetooth radio and a WLAN radio source have been considered. in fact the problem can be interpreted as “how to observe the world”. 5. and how this usage can improve the performances of the set of CTs. a sensing approach based on Wigner-Ville (WV) time-frequency analysis [16]. Proposed Framework The proposed framework is an similar to the one proposed in [15]. by exploiting these differences. come. where. two features are extracted: • the standard deviation of the instantaneous frequency. similar to the one presented in [15] is presented: W (t . is to observe the radio channel and to decide the current transmission mode. finally Decision and Action modules. • Both WLAN and Bluetooth sources are transmitting. for minimizing memory occupancy and exchange easiness. This information is used in the online phase to allow the two CCRs to decide the current transmission situation in a distributed way. Information acquisition: Sensing Map and procedures In the present paper. Figure 3 Instantaneous Frequency for WLAN and Bluetooth Figure 2 Proposed System Architecture The first one strictly related to the Sensing Map. The Cognitive Maps extraction process can be considered an acquisition and formalization of a-priori knowledge about the environment and it is performed in an off-line phase. In Figure 2 the proposed system architecture is shown: after a first analog frequency down-conversion to intermediate frequency stage. PROPOSED APPROACH The proposed approach deals the problem of information acquisition and handling for cooperative CTs. from the WV transform. at the same time. and it can be divided into three sub-problems: • Information acquisition • Information representation and storing • Information usage Once the information has been expressed in a synthetic form. • the maximum time duration. it’s possible to define a uniform Sensing Map: in fact over all the room information about the radio channel will be acquired through WV time frequency transform and its processing. Two Cooperative Cognitive Radios (CCRs) explore the environment trying do decide the current radio context in a distributed way. completely understandable by the CTs. in a 2D space. in every point of the room. The room is explored by two cooperative CTs. 4. respect to the radio environment. because its solution should respect some constrains: the representation should be synthetic.1. an analog-to-digital converter is inserted. together with the time-frequency transform. The second one is more complex. the Analysis stage.

an architecture which allows to extend the Varshney’s theory to a multi-class problem has been designed: Figure 5 Analysis Map for Bluetooth Class obtained in a simulated environment Proceeding of the SDR 06 Technical Conference and Product Exposition.5. by knowing the position of the sensor.i ≠ k where Pi and Pk are the prior probabilities and Bik is the Bhattacharyya distance between i and k. the upper bound for the probability of confusing k with another class can be written as: This representation of the information acquired in the off-line phase has two main advantages: the first one is that it’s a complete representation of the classification behavior of a stand-alone CT immersed in the environment. In the present paper. This simplification. it’s possible to proof that the only presence of another detector inside the environment influence the decision processes. Information representation and storing: off-line data acquisition and Analysis Map creation In an off-line phase. a variation of the distributed detection proposed by Varshney [5] has been considered. Give the presence of multiple classes in the considered framework. as a 2D Gaussian in the features space. it’s possible to store only the interpolation coefficients which generate the surface.2. The decision framework is the so-called Distributed Bayesian Detection without Fusion. The interpolation has been performed through a simple cubic 2D polynomial. k two classes. thanks to the relationship between the distance and the Chernoff bound [18]: being i. is related to the terminal labeled as “1”: it’s possible to define similar test and threshold for terminal “2”. how to use it in the best way has to be decided. in every transmission situation. features have been extracted from different realizations of the radio signal. The right-hand side of the equation is the classical Bayesian likelihood function. while the left-hand side is the distributed threshold which can be re-written as: Figure 4 Analysis Map for WLAN Class obtained in a simulated environment The threshold. 5. a 2D extension of [15]. Starting from the detection of a binary phenomenon. All Rights Reserved . decreases the analytic complexity of the problem and it allows the usage of an inter-class distance measure.3. A “One shot” parameter estimation has been used in order to estimate the polynomial coefficients. Copyright © 2006 SDR Forum. Thanks to the acquired features. in each point. The second one is that. the obtained values are interpolated to obtain a surface which represent the goodness of the classification process in each point of the room. such as the Bayesian test. It’s hence possible to obtain an upper bound for the error probability for each considered class. Information usage: Distributed Classification and Maps modification Once acquired the information. It is substantially based on a Bayesian likelihood function compared with a threshold that depends on the position of the other CCR and which is its point-of-view about the radio world in the considered point fo the environment: Pek ≤ 1 4 ∑ Pi Pk exp(− Bik ) 3 i =1. it’s possible to model the probability density function (pdf) of each class. the Bhattacharyya distance [17]. The upper bound of the error probability for each class in each sample point of the room is used to obtain the Analysis Map of the environment: in order to provide a continuous map over all the room. in different points of the room.

in a binary way. 5. one for each class present in the problem. For K = 3 and 5 more complex surfaces are obtained. as described by Varshney’s theory. Being the decision thresholds function of the error probability of the class. x2 ) + P1 p (C 2 m = 1 | H 1 . one has a remarkable error frequency (Wlan. to approximate the distributed threshold: in fact it is considered as a function of the error probability for the class under test: t1m 1 P0 + ( K − 2) Pem ( x 2 ) − ( K − 2) P1 Pem ( x 2 ) 8 = 1 ( K − 2) P1 Pem ( x 2 ) + 1 − P0 8 Mean Error Frequency Table 1 Error Frequency in the stand-alone mode (K = 2) where m represents the tested class. In fact each classifier tests. Increasing K means to give more importance to the considered class in the comparison with the likelihood function. x2 ) P1 ( K − 1) p (C 2 m = 0 | H 1 . The terminal is also aware of its position and of the position of the companion terminal.Figure 6 Multi-Class distributed architecture Figure 8 Modified Analysis Map for Bluetooth Class obtained in a simulated environment Four classifiers have been considered. i. while x2 represents here the position of the companion terminal. while Proceeding of the SDR 06 Technical Conference and Product Exposition. K=5 Real Class Wlan Bluetooth Wlan + Bluetooth Noise Mean Error Frequency x1 = (11. 40%) while Bluetooth is never identified in a correct way. It represents the stand-alone classification for the considered class. P0 and P1 are the prior of the two possibilities (presence or absence of the class). thanks to the previously obtained Analysis Maps. in a distributed way. that has a substantial reduction of the classification error. 5. For a given couple of positions of the two terminals. one class against the others. But it’s also possible to see that this advantage is clear for one terminal only. x2 ) + P0 ( K − 1) p (C 2 m = 1 | H 0 . it’s hence possible to derive a distributed threshold for each classifier.e. the following classification results have been obtained: K=2 Real Class Wlan Bluetooth Wlan + Bluetooth Noise x1 = (11.5. it’s possible to derive new Analysis Maps which represent the variations of the threshold over all the room. i. K is a decisional weight. x2 ) In Figure 7 and Figure 8 three different surfaces for each class are shown: for K = 2 a flat surface is obtained.11) Error Frequency 40% 0% 0% 0% 10% x2 = (3.e. Copyright © 2006 SDR Forum. All Rights Reserved .11) Error Frequency 40% 100% 0% 0% 35% x2 = (3. Different surfaces for different values of K can hence be obtained: In the considered points of the environment some classes are correctly identified (Wlan+Bluetooth and Noise).5. for each class under test: t1m = P0 p (C 2 m = 0 | H 0 . it’s possible. and being the error probability function of the position of the terminal.3) Error Frequency 40% 70% 90% 0% 50% Table 2 Error Freqeuncy in distributed classification mode (K = 5) Figure 7 Modified Analysis Map for WLAN Class obtained in a simulated environment In the same positions it’s possible to see how distributed detection mode is more efficent than the stand alone mode.3) Error Frequency 20% 100% 100% 0% 55% Being the classification probability distribution for terminal “2” considered unknown.

Oct. Proceedings of IEEE. Springer-Verlag. vol. Urkowitz. vol. Regazzoni. are shown. 1. [13] P. no.F.Willett. For a specific problem. pp. April 1967. chapter in Software Defined Radio.M. no. IEEE Journal Selected Areas in Communication. Cohen.. the Distributed Mode Identification problem. 1996 [6] H. pp. Cattoni. Varshney. no. Rago and Y. Chair. Roland. REFERENCES [1] J. Notice of proposed rule making and order. vol. Cognitive radio: making software radio more personal. Cognitive radio shows great promise. International Journal of Distributed Sensor Networks. Time Frequency Analysis : Theory and Applications. Faroughi-Esfahani. which is completely understandable by CTs and it’s useful for their target reaching. Gandetto. Prentice Hall PTR. August 1999. Special Issue on Non Linear Signal Processing and Image Processing. Vardoulias and J. Results proof that it’s possible to obtain a synthetic representation of the physical world. 554-568.K. On a measure of divergence between two statistical populations defined by probability distributions. [5] P. [17] A. Chernoff. 98101. [4] Federal Communications Commission. no. Cognitive radio: brain-empoweres wireless communications. 55. 23. Math. John Wiley and Sons Ltd. 523531. pp. 7. together with the importance that the so-called Cognitive Maps have in the considered model. available at http://www. ET Docket 03-322. 1st edition. 1778-1790. 6. pp. 54-63. [7] J. C. no. vol. 23. Varshney.the other CCR has a very few improvement in classification. 40. Bhattacharyya. Proceeding of the SDR 06 Technical Conference and Product Exposition. July 2003. these motions can be interpreted by the other terminals exploting the characteristics of the cognitive modeling) way. 4. April 2003. S. [9] M. pp. 41. 13. is useful to improve the classification performances of a set of CCRs composed by two terminals. C. the development of also Decision and Action stages is subject of on-going researches. Mode Identification and Monitoring of Available Air Interfaces. Ann. Viswanathan. no. Energyeffcient detection in sensor network. 2. Tech. 2004. 22. In order to complete the cycle. Fette. vol. V. Bar-Shalom. Palicot. no. Appadwedula. pp. Optimum data fusion in multiple sensor detection systems. 1943. [15] M. Varshney.S. IEEE Transaction on Communications. Architectures. A. 85.K. [10] C. Censoring sensors: A low-communication-rate scheme for distributed detection. 201-220. [11] P. In the present paper the design of only Sensing and Analysis stages has been considered. Signal interception: performance advantages of cyclic-feature detectors. vol. A Distributed Wireless Sensor Network for Radio Scene Analysis. 1. A measure of asymptotic efficiency for tests based on the sum of observations. December 1994. 124-132. After a brief overview on how their concept has been defined by international researchers. [8] G. April 2005. Eurasip Journal of Applied Signal Processing. 1. a behavioral model for CTs. on Aerospace and Electronic System. W. April 1996. obtained in a completely simulated framework. Stat. 6. no. 99-109. This information will be transmitted to the other terminal in an explicit (peer-topeer or broadcast radio communication) or implicit (special motions of the CCR in the environment. 149-159. Distributed Detection and Data Fusion. pp. System and Functions. no. on Aerospace and Electronic System.Distributed detection without fusion. 4. is provided. 1st edition. vol. January 1997.K. 32. FCC. vol. Haykin. pp. 2. R. Feb6braio 2005. The representation respects also the requirement to be extremely compact (only interpolation coefficients are stored). Regazzoni. 2006 [16] L. All Rights Reserved . extended to a multi-class decision problem.A. Gandetto. Z. December 2003. 1952. it’s sufficent that at least one of them is able to perform a good classification. January 1986. IEEE Communications Magazine.fundamentals. IEEE Trans. Finally classification results obtained in a completly simulated framework are presented: they proof that distributed detection theory. [2] B. Use of time-frequency analysis and neural networks formode identification in a wireless software-defined radio approach. A new concept for wireless reconfigurable receivers. CONCLUSIONS This paper try to face the problem of information representation and usage in Cognitive Radios. IEEE Trans. Gardner. 693-702. chapter 3 . Energy detection of unknown deterministic signals.L. IEEE Selected Area in Communication. Veeravalli.cotsjournalonline. IEEE Personal Communication. Taylor & Francis Publishing.e.329-352. i. Bullettin of Calcutta Mathematical Society. Prentice-Hall Signal Processing. C. [12] P. Distributed detection with multiple sensors: Part I . [18] H. a possible extraction process is pointed out and some Cognitive Maps. [3] S. Copyright © 2006 SDR Forum.com/. vol. pp. Guainazzo and C. by considering all the CCRs in the environment a team of classifier. pp. pp. Jones. pp. Spooner. Proceedings of the IEEE. 48-52. vol. Mitola. COTS Journal. pp. 7. These results require a specific point of view in order to completely understand the real significance of distributed detection: in fact. 4. [14] D. January 1992. Rep. 493-507. 2005. M. S.

Sign up to vote on this title
UsefulNot useful